memtx: accelerate count in fast_offset tree

This patch creates a specialized implementation for `count` tree index
method, which takes advantage of offset-based functions of the BPS tree
with BPS_INNER_CARD config.

In order to handle keys and iterators consistently with `select`, the
iterator creation and first step preparation algorithms had been moved
into separated functions.

NO_DOC=see the next commit
NO_CHANGELOG=see the next commit